wpc0323/deepseek-web-import

wpc0323/deepseek-web-import

从设置页把 chat.deepseek.com 网页端对话导入为 DeepSeek Harness 会话,导入后可在其中继续对话。

会话与消息

DeepSeek Web Chat Import (DeepSeek Harness plugin)

English | 中文

Import your chat.deepseek.com conversation history as real DeepSeek Harness (DSH) sessions: each import appears in the sidebar workspace list, opens with its full history, and can be continued with any model.

This is an installable DSH plugin bundle (dsh.bundle manifest: a host half plus a web client half), installed with dsh plugin add. See Install below.


Screenshots

The DeepSeek Chat Import settings page: sign-in state, conversation directory, and the collapsed diagnostics section


✨ Features

  • A "DeepSeek Chat Import" page inside the settings panel
  • Paste a userToken and fetch the conversation directory in one click (titles + dates only; transcripts are fetched on import)
  • Pick a target workspace per conversation and import it as a DSH session in one click
  • Imported sessions:
    • keep the original DeepSeek title
    • can be continued with a model of your choice (they take DSH's normal resume path, so they never hit cannot prepare session while it is live)
    • are persisted under ~/.dsh/sessions/<workspace>/ and survive a DSH restart
  • Extracts the value field from a userToken automatically (accepts the {"value":"...","__version":"0"} form)
  • A built-in Diagnostics section for requesting raw responses when the upstream API changes
  • Localized UI (English/Chinese) that follows the DSH/browser language and falls back to English

⚠️ Important

Refresh the page (F5) after importing a conversation so the sidebar shows the new session immediately.

Without a refresh the new session may only appear after the next session-list refresh or reconnect.


Install

Prerequisite: a node executable on PATH (Node.js). The plugin performs its request in a node child process (DSH's web.fetch cannot carry custom headers); without Node.js the page says it cannot start Node.js, and why, instead of failing to load the directory for no visible reason.

The plugin declares dsh.bundle (see package.json), so it installs with dsh plugin add:

dsh plugin add deepseek-web-import        # published on npm
# or straight from this repository:
dsh plugin add wpc0323/deepseek-web-import

After installing, restart DSH; Settings → DeepSeek Chat Import then appears. In this repository lib/index.js is the host half (same-origin JSON routes) and lib/client.js is the web client half (the settings-page UI); they talk over the webServer routes and fetch.


Usage

  1. Open Settings → DeepSeek Chat Import.
  2. ① Sign-in: paste the userToken → Save.
    • How to get it: sign in to chat.deepseek.com in your browser → F12 → Application → Local Storage → copy the userToken value.
    • The plugin extracts the value field itself, so pasting the whole entry is fine.
  3. ② Conversation directory: click "Load conversations" → the most recent 100 conversations appear (title + date).
  4. For each conversation: pick a workspace, then click "Import as session".
  5. Refresh the page after importing, then open the new session from the sidebar → choose a model → keep chatting.

Example

1. Open Settings → DeepSeek Chat Import
2. Paste the userToken → Save (the status turns "configured")
3. Click "Load conversations" → your DeepSeek conversations appear (title + date)
4. On the "sharp load failure fix" row, pick the workspace dsh-work → click "Import as session"
5. Refresh the page → "sharp load failure fix" appears in the sidebar → open it and continue the conversation

An import writes to persistence only (sessionPersistence) and never enters the live store, which is why an imported session can be opened and continued normally (it never reports cannot prepare session while it is live).


A session will not open, or the log looks corrupt

See docs/TROUBLESHOOTING.md (Chinese). It covers two failures:

  • corrupt session log: seq gap … — a DSH recovery-cursor bug of its own (session/end-seed is appended separately by a live session, and the post-restart cursor misses it, so a seq gets reused); unrelated to this plugin, whose bulk write happens to avoid the trigger.

  • format v2 surface before first step … — a session imported by plugin 0.1.0 cannot be migrated by a v3+ DSH build, because that importer wrote turn/start → user/message → step/start while the v2→v3 migration requires no surface event before the first step/start. 0.2.x writes the correct order, so new imports are unaffected; an old artifact can be repaired with the bundled tool:

    node tools/fix-surface-before-step.mjs --root ~/.dsh/sessions          # dry run: what would change
    node tools/fix-surface-before-step.mjs --root ~/.dsh/sessions --apply  # back up, then repair in place
    

Compatibility (DSH Session format v0–v4)

Every DSH release writes one session format (v0…v4). The plugin adapts to the running build. Measured matrix:

DSH versionSession formatStatus
0.1.0-rc.7 – 0.1.2-rc.1v0✅ verified against the real backend
(unreleased)v1✅ modeled as v0 + runtime dialect evidence
0.1.3-alpha.2v2✅ verified against the real backend
0.1.5-alpha.1 – 0.1.6-alpha.2v3✅ verified against the real backend
0.1.7-alpha.1+ (including desktop 0.2.0-rc.2)v4✅ verified against the real backend

How each row was verified, and where the evidence comes from, is in docs/COMPATIBILITY.md (Chinese). npm test runs the unit tests; sh test/matrix.sh runs the full matrix.

How it works (briefly)

  • Architecture: lib/index.js (host half) registers same-origin JSON routes (/__deepseek-web-import/*) on DSH's webServer; lib/client.js (web client half) calls them with fetch from the settings page. No dynamic-plugin machinery is involved, so the plugin stays installed across restarts.
  • Endpoints (unofficial DeepSeek web internals, confirmed by reverse-engineering their own JS bundle):
    • conversation directory: GET /api/v0/chat_session/fetch_page?count=100
    • conversation history: GET /api/v0/chat/history_messages?chat_session_id=<uuid>
  • Import as session: history messages are translated into valid DSH session events (turn/start → user/message → step/start → assistant/message → step/end → turn/end, surface events carrying surfaceOp: "append", plus session/title and session/end-seed), written to persistence only (v0/v1 through create(meta) + append(id, events), v2+ through the write handle returned by create(header)), and then attached to the chosen workspace with workspace.attachSession — the session never enters the live store, which is what makes it openable and continuable. A history that ends with an unanswered user question keeps its unclosed turn; DSH closes it with interrupted when the session is opened.
  • Body and reasoning: DeepSeek keeps the text in fragments (REQUEST/RESPONSE/THINK). A user question comes from REQUEST; an answer uses RESPONSE as the text and THINK as a reasoning block. Event times come from the message's inserted_at.
  • Network layer: web.fetch cannot carry custom headers, so the plugin runs Node through the subprocess service and sends the request with Authorization: Bearer. The request spec (including that header) travels on the child's stdin, so it never shows up in ps//proc. Certificates are verified first; only a certificate-class failure (an environment whose TLS is intercepted by a proxy or antivirus) retries once without verification and reports insecureTls (one line in the DSH log, visible in Diagnostics) instead of degrading silently. The child bounds the streamed response by characters (~2M) and by bytes (6MB) and reports too_large instead of buffering a huge response into memory.
  • Interface language: settings-page copy is registered into the host locale namespace (ctx.locale), so it follows the DSH/browser language and falls back to English for unregistered languages (German, say). Without a locale service the module resolves the browser language itself.
  • Token storage: through DSH's credentials service ($DSH_HOME/.credentials.yaml); it is never written into a session or the settings document.

Known limitations

  • It depends on DeepSeek's unofficial internal API, which can break when the web app is redesigned (use Diagnostics to see the raw response when it does).
  • Password sign-in is blocked by DeepSeek's bot protection (AWS WAF), so the userToken route is the only one.
  • The directory fetches at most 100 conversations: the measured fetch_page response carries only chat_sessions + has_more, with no usable cursor field (cursor/lte_cursor have no effect), so beyond 100 you get the latest 100.
  • userToken expires after some hours to days; copy and save a fresh one when it does.
  • The plugin installs as a bundle through dsh plugin add and stays installed across DSH restarts.
  • Web-search steps (TOOL_SEARCH/TOOL_OPEN fragments) and message status (an INCOMPLETE answer cut off while generating) are not preserved: web search has no DSH event to map onto, so only the text and reasoning are imported; an interrupted answer keeps whatever text/reasoning it had.
  • Importing the same DeepSeek conversation twice creates two DSH sessions (the plugin keeps no "DeepSeek conversation id → DSH session id" mapping) and DSH's persistence offers no delete, so archive the duplicate yourself.

  • The response cap for one import is about 2M characters (a conversation over it returns too_large rather than being silently truncated).
  • An attachment-only message imports as one explicit note; non-conversation roles (SYSTEM/TOOL, …) are ignored.
  • An import follows the single main line returned by history_messages (it does not resolve parent_id branches): if DeepSeek returns sibling branches for an edit/regenerate, they are imported as two consecutive turns.

Security

  • Never leak your account password or userToken in a conversation, issue, or repository.
  • If you have pasted your password into a chat, change your DeepSeek password now.

License

MIT